Disclosed is an effective algorithm to correct motion-induced phase error using an iterative reconstruction. Using a conjugate-gradient (CG) algorithm, the phase error is treated as an image encoding function. Given the complex perturbation terms, diffusion-weighted images can be reconstructed using an augmented sensitivity map. The mathematical formulation and image reconstruction procedures are similar to the SENSE reconstruction. By defining a dynamic composition sensitivity, the CG phase correction method can be conveniently incorporated with SENSE reconstruction for the application of multi-shot SENSE DWI. Effective phase correction and multi-shot SENSE DWI (R=1 to 3) are demonstrated on both simulated and in vivo data acquired with PROPELLER and SNAILS.
|
1. A method of magnetic resonance imaging reconstruction with phase error correction comprising the steps of:
a) Obtaining data for a plurality of k-space trajectories using a plurality of receiver coils, each with a specific coil phase,
b) Defining data in terms of coil sensitivity and position vectors,
c) Modifying the coil sensitivities for each coil in step (b) to include a time invariant and time variant phase perturbation caused by (i) motion during diffusion encoding, and (ii) at off resonance blurring, and
d) Iteratively reconstructing the data to form image data for image reconstruction by employing an iterative conjugate gradient reconstruction.
18. A method of magnetic resonance imaging reconstruction with phase error correction comprising the steps of:
a. Obtaining data for a plurality of k-space trajectories using a plurality of receiver coils, each with a specific coil magnitude and phase,
b. Defining data in terms of coil sensitivity and position vectors,
c. Modifying the coil sensitivities for each coil in step (b) to include a time variant spatially variable phase perturbation caused by (i) motion during diffusion encoding, which varies between interleaves, blades, and k-space lines following each RF excitation or diffusion step, and (ii) spins at off-resonance, which continuously accrue a phase angle during each interface, blade, and k-line, and
d. Iteratively reconstructing the data to form image data for image reconstruction by employing an iterative conjugate gradient reconstruction or similar algorithms for large matrix systems.
2. The method, as defined by
reformulating a reconstruction problem into: (EHE)v=EHm , where
v is a voxel matrix, E is an image encoding matrix comprising the modified coil sensitivities and a discrete Fourier Transform, and m is a vector containing the acquired k-space sampling data;
precomputing EH m once;
initialize v(0); and
solving for v iteratively by computing (EHE)v(n), where n indicates an iteration step, so that v converges to a solution that minimizes a norm∥(EHE)v(n)−EHm∥.
3. The method as defined by
4. The method as defined by
where
sγ=coil sensitivity
r(ρ)=position vector
V=voxel
e−jk
6. The method as defined by
7. The method as defined by
8. The method of
9. The method as defined by
10. The method as defined by
11. The method as defined in
12. The method, as defined by
reformulating a reconstruction problem into: (EHE)v=EHm, where
v is a voxel matix, E is an image encoding matrix comprising the modified coil sensitivities and a discrete Fourier Transform, and m is a vector containing the acquired k-space sampling data;
precomputing EHm once;
initialize v(0); and
solving for v iteratively by computing (EHE)v(n), where n indicates an iteration step, so that v converges to a solution that minimizes a norm ∥(EHE)v(n)−EHm∥.
where
sγ=coil sensitivity
o(rρ(iΔt))=off-resonance phase
r(ρ)=position vector
v=voxel.
14. The method as defined by
15. The method as defined by
16. The method of
17. The method as defined by
|
The U.S. government has rights in the claimed invention pursuant to NIH Grant No. 1R01EB002711 to Stanford University.
This invention relates generally to multi-dimensional magnetic resonance imaging (MRI), and more particularly the invention relates to MRI reconstruction in the presence of multiple image signal distortions.
Magnetic resonance imaging (MRI) is a non-destructive method for the analysis of materials and represents a new approach to medical imaging. It is generally non-invasive and does not involve ionizing radiation. In very general terms, nuclear magnetic moments are excited at specific spin precession frequencies which are proportional to the local magnetic field. The radio-frequency signals resulting from the precession of these spins are received using pickup coils. By manipulating the magnetic fields, an array of signals is provided representing different regions of the volume. These are combined to produce a volumetric image of the nuclear spin density of the body.
MRI signals for reconstructing an image of an object are obtained by placing the object in a magnetic field, applying magnetic gradients for slice selection, applying a magnetic excitation pulse to tilt nuclei spins in the desired slice, and then detecting MRI signals emitted from the tilted nuclei spins while applying readout gradients. The detected signals can be envisioned as traversing lines in a Fourier transformed space (k-space) with the lines aligned and spaced parallel in Cartesian trajectories or emanating from the origin of k-space in spiral trajectories.
Self-navigated interleaved spirals (SNAILS) have been used for high resolution diffusion weighted imaging (DWI). Although spiral trajectories have many advantages for fast image acquisition, they normally suffer from image blurring caused by off-resonant spins. Many techniques have been developed for off-resonance correction. However, few studies have been reported for off resonance correction for multi-shot DWI. One difficulty in this situation originates from the k-space data distortion caused by motion-induced phase errors.
Diffusion-weighted imaging (DWI) provides a unique tissue contrast by sensitizing random molecular thermal motion using magnetic field gradients. Because of its ability to quantify this random motion, DWI has become a powerful tool for studying tissue micro-structures and detecting acute ischemic stroke in which diffusion is highly restricted very early after the onset of stroke.
Patient motion during diffusion encoding results in additional phase terms that lead to severe ghosting artifacts if not accounted for. To avoid such artifacts, diffusion-weighted images are therefore most commonly acquired using single-shot echo planar imaging (EPI) sequences. The drawbacks of single-shot EPI include relatively low image resolution limited by T2* decay and geometric distortions caused by magnetic field inhomogeneity. One technique to overcome the aforementioned distortions is combining sensitivity encoding (SENSE) with single-shot EPI. Alternatively, several multi-shot sequences, for example PROPELLER and SNAILS (self-navigated interleaved spiral), have been shown to be capable of acquiring high resolution (up to 512×512) diffusion-weighted images with significantly diminished distortions.
One common problem inherent to multi-shot diffusion-weighted image acquisition is image-domain phase perturbation that varies from shot to shot, which has thus far limited the application of parallel imaging to multi-shot DWI. This phase variation can be either linear or non-linear. The linear phase variation is usually caused by rigid-body motion during diffusion encoding periods; whereas the non-linear phase can be caused by nonrigid motion, for example by brain pulsation. Correcting this phase variation by subtracting a low resolution phase map from each shot has proven efficient. The phase map can be obtained either from an extra navigator image or from the k-space data of a self-navigated trajectory (e.g. variable density spirals). This simple phase correction algorithm can remove effectively the phase error to a certain degree which has been applied in both PROPELLER and SNAILS DWI. This algorithm was thought to be adequate, provided that each shot samples the k-space above the Nyquist rate. PROPELLER DWI is one technique that satisfies this sampling criterion. More specifically, in PROPELLER DWI, during each shot a selected segment of the k-space is sampled at the Nyquist rate. Data acquired from each shot results in a unaliased, but blurred, image, which permits a straight-forward phase subtraction.
Although successful image reconstruction has been demonstrated with PROPELLER DWI, the phase variation cannot be completely corrected by the direct subtraction algorithm due to image blurring or aliasing. For PROPELLER DWI, because the phase error is usually smooth, the residual artifact might not be so severe. For arbitrary, undersampled k-space trajectories, however, the incomplete phase correction worsens considerably because of possible aliasing artifacts. For example, when each shot undersamples some portions of the k-space, the effect of aliasing causes the phase error at one location to appear at other locations. The resulting non-localized phase error can no longer be corrected through a simple phase subtraction. Therefore, images reconstructed using this simple phase subtraction algorithm may still suffer from severe motion-induced artifacts and residual ghosts may have to be suppressed using multiple averages.
The present invention is directed to a generalized iterative image reconstruction which simultaneously corrects for multiple signal distortions including phase, off resonance and gradient non-linearities.
The present invention acquires two- and three-dimensional magnetic resonance signals and performs a generalized iterative image reconstruction including correction of phase, off resonance, and gradient non-linearities.
In a preferred embodiment, the invention utilizes an iterative conjugate gradient approach but is applicable to eventual faster algorithms for matrix inversion, or solving the entire reconstruction problem by means of Calculus of Variation (Total Variation). Moreover, due to the nature of the iterative reconstruction, no a priori knowledge of the sampling density for retrospective correction is necessary. While prior art can suffer from strong geometric distortions and blurring or ghosting errors, this invention is unique since it corrects for all these distortions simultaneously.
Although not limited to, the method capitalizes on variable-density spiral acquisitions from which navigator phase and complex coil sensitivity information can be acquired. Alternatively, this information can be obtained from additional navigator echoes and calibration scans addressed in prior art. Off-resonance phase information can be obtained by slightly shifting the echo time between two acquisitions, also addressed in prior art.
The invention permits correction of all of these distortions simultaneously by incorporating the perturbations directly into a design matrix of the image reconstruction formalism.
The invention and objects and features thereof will be more readily apparent from the following description and appended claims when taken with the drawings.
Consider first the conventional, generalized discrete notation for MRI data acquisition for an arbitrary k-space sampling pattern.
The signal formation in k-space for a multi-coil MR experiment with an arbitrary k-space trajectory can be written for the γth-coil as:
or more explicitly written for the γth-coil as:
Here, the vectors m=[m0 m1 . . . mnc]T and v are of lenthl nk*nc and N*N, respectively (nk denotes the total number of k-space sampling points, nc denotes the total number of receiving coils, and N denotes the number of points along each dimension of the image), and consequently the matrix E=[E0 E1 . . . Enc]T and (EHE) must be of size nk*nc×N*N and N*N×N*N, respectively. This is basically the discrete Fourier Transform in matrix notation modulated by the complex coil sensitivity s(r) that—applying the principle of reciprocity—each voxel r sees. Notice that the image signal is stored in vector form v. The vector indices in Equation 2 are suggesting a 2-dimensional reconstruction. However, due to the vector notion this problem can be easily extended to 3-dimensions. Increased reconstruction time should not be our current concern for the proof of concept. The only requirement to solve Equation 2 is nk*nc≧N2. In a typical spiral imaging experiment this condition is generally fulfilled given that every 4 us a data point is generated and the trajectory design fulfills the Nyquist sampling criterion. As long as this problem is sufficiently conditioned and not underdermined, an image can be reconstructed without knowledge of the degree of undersampling. This reconstruction is also applicable for sufficiently sampled single-coil images without knowledge of the coil sensitivity in practical applications, the DFTs will be replaced by (gridding and FFT/IFFT (Fast Fourier Transform and Inverse Fast Fourier Transform) and ultimately by convolutions.
For N>128, clearly this inverse problem is excessively large for direct computation and other methods should be considered. By observing that m=Ev minimizes the quadratic form1
f=0.5vTEv−mTv+c (3)
the generalized SENSE (GSENSE) reconstruction problem can be seen as finding the solution vector v such that the quadratic form has its minimum. The simplest way of finding this solution would be to iteratively follow the steeliest gradient of f. Unfortunately, this method converges very slowly if E is poorly conditioned and an unfortunate starting location is chosen. A very effective method to solve the (generalized SENSE reconstruction problem is the conjugate gradient (CG) method1:
1. reformulate the reconstruction problem into:
(EHE)v=EHm (4)
2. precompute EHm once (gridding).
3. initialize v(0) with zeroes.
4. solve for v iteratively by computing (EHE)v(n), where n indicates the iteration step.
The expression (EHE)v(n) can be effectively solved by two gridding operations, where k=Ev(n) is an inverse gridding operation2: i.e., (i) perform roll-off correction, (ii) zero-pad v(n) and inverse FFT, (iii) convolve with Kaiser-Bessel window and resample k-space data along the prescribed k-space trajectory. Likewise, the (EHE)v(n)=EHk is a conventional gridding operation: i.e., (i) convolve k-space trajectory data with Kaiser-Bessel window and resample on a 2× grid, (ii) FFT to image domain and trim image matrix, (iii) perform roll-off correction.3 A nice feature of this two-step process is that the matrix E is not required to be precomputed or stored. Moreover, utilizing the FFT reduces the computational complexity of each step from N4 to N2log N operations.
The only difference from the classical gridding operation proposed by Jackson el al. is the missing sampling density correction. However, this can be accounted for by introducing a diagonal matrix D (D(γ,K),(γ,K)=k_density(kK)−1) into the reconstruction problem (EHDE)v(n)=EHDk. This step leads to a much faster convergence of the CG algorithm. To further improve convergence, an image intensity correction should be performed for the right-hand side of Eq. 4. The right-hand side expression of Eq. 4 is weighted by the sum-of-squares of the coil sensitivities and should thus be corrected by using the root-mean-square of the intensity weighting factor
Cρ,ρ=1/√{square root over (|sγ(rρ)|2)}. (5)
However, left-multiplication of the left-hand side of Eq. 4 means there is no longer a guarantee of positive definiteness (i.e., vTEv>0 for all v≠0) and, therefore, both a left-multiplication with C and the introduction of the identity matrix CC−1 is performed, which still improves convergence of the conjugate gradient method:
(CEHDEC)(C−1v)=CEHDm. (6)
Here, the right-hand side expression represents the gridded, density- and intensity-corrected undersampled image. An intensity modulated image (C−1v) rather than v will be evaluated during each iteration step.
Since virtual receiver channels without mutual correlation can by generated from the true channels by means of Cholesky decomposition, in the following, the noise statistics should be neglected, and the generalized SENSE reconstruction problem can be expressed as
EHEx=EHm. (7)
As already discussed, this problem can be solved effectively by means of the conjugate gradient algorithm. The right-hand side of this equation needs to be evaluated only once, whereas the left-hand side will be iterated until the estimate x(n) converges to the true solution. In the original GSENSE reconstruction EHEx(n) is evaluated in two steps as EH(Ex(n)), which can be effectively accomplished by forward and inverse gridding operations. These two gridding steps heavily determine the calculation time during each iteration and Wajer et al. showed that a reorganization of this expression can dramatically accelerate the reconstruction time. By examining the components of the matrix product EHE, it becomes immediately evident that its elements are:
The expression for Q(r) can be precomputed together with EHm prior to starting the conjugate gradient iterations. The only difference is that the size of Q is twice as large as x(n) since the argument of Q(r), rρ−rρ′, has twice the range of rρ. For a particular voxel position rρ, the elements of EHEx(n) can then be evaluated by a discrete convolution of Q with (sγx(n)
and can be efficiently solved using fast fourier transform (FFT), i.e.
└Q*=(sγ(rρ′)xp′(n))┘=IFFT{FFT[Q]·FFT└sγ(rρ′)xp′(n)┘}. (11)
By storing FFT[Q] instead of Q only two FFTs are required, which have the size of Q and require sγ(rρ′)xp′(n) to be zero-padded prior to each iteration step. After this convolution the result has to be multiplied by the complex conjugate of the coil sensitivity sγ(rρ). . . and summed over all nc coils. The net increase in computational speed results from replacing the forward and inverse gridding interpolation2 by a coil-wise multiplication in k-space with FFT[Q], which is much faster than the gridding operations. Finally, the k-space density correction2 can be implemented without any further problems by a slight modification of
The algorithm described above was implemented in IDL (Interactive Data Language) on 2.0 GHz Pentium4 notebook and serves as framework for our GSENSE reconstruction. For a single-shot spiral acquisition (128×128 matrix size, 2× grid, 6-element array) the precomputation for EHm and Q took 3.48 s and 2.58 s, respectively. Each conjugate gradient iteration step took 0.95 s. Likewise, for a single-shot spiral (256×256 matrix size, 2× grid, 6-element array) the precomputation for EHm and Q took 13.97 s and 10.68 s, respectively. Each conjugate gradient iteration step took 8.55 s. With the same method, the convergence for different degrees of SENSE reduction of a single-shot spiral (128×128 matrix size, 2× grid) was analyzed.
Equation 1 can be modified so that the measured magnetization considers time invariant phase perturbations as Such from motion during diffusion and time-varianit phase perturbations as such from off resonant spins. First, we will include time invariant phase errors. Typically, such phase errors are phase perturbations in the image domain.
i) Time Invariant Phase:
Hence, modifying Eq. 1 will lead to:
where p(rρ) is the time invariant phase perturbation (for example for diffusion) at the particular voxel of interest. This perturbation can be easily incorporated into the existing iterative reconstruction by means of an augmented sensitivity that includes the perturbation phase in addition to the coil specific B1 phase, i.e.
{tilde over (s)}γ(rρ)=sγ(rρ)p(rρ) (14)
and replaces the original sensitivity information.
ii) Time Variant Phase:
Another modification Eq. 1 will consider time variant phase due to off resonant spins. Given that the spatial distribution of off resonance is known and considering that k is linear in time Eq. 1 can be rewritten as
where o(rρ(iΔt) is the time variant phase perturbation due to off-resonant spins at the particular voxel of interest, e.g.
o(rρ)(iΔt))=exp(j2πΔf(iΔt)) (16)
This perturbation can be easily incorporated into the existing iterative reconstruction by means of an augmented sensitivity that includes the perturbation phase in addition to the modifications made in Eq. 14, i.e.
{tilde over (s)}γ(rρ(iΔt))=sγ(rρ)p(rρ)o(rρ(iΔt)). (17)
Since all elements of matrix E have to computed anyway it is no penalty to have s time varying, i.e.
Providing that the gradient non-uniformity for each of the gradient coils is known at each point within the prescribed scanning slab, the actual k-vector can be expressed as
Here, the vector G(t)=[Gx(t)Gy(t)Gz(t)]T is the desired (spatially uniform) gradient waveform. If the relative deviations from the desired field are known, a spatially varying gradient coil tensor L(r) can be used to relate the actual gradient produced by the coil to the desired gradient. The elements of the gradient coil tensor contain the spatially varying error terms for each of the principal gradient axes, where the actual gradient, Gact, and the desired gradient, G, are related by:
where L(r) is the gradient coil tensor. In equation 20, matrix element Lij(i,j=x,y,z) represents the i-component of the magnetic field gradient when a unit gradient in direction j is intended. The Lij can be calculated from the partial derivatives of the effective spatially dependent field normalized by the applied nominal gradient strength Gi
Here, the spatial dependence of the field B=i (r) generated by each gradient coil (i=x, y, or z coil) can be generally expressed as a spherical harmonic expansion8
where HklC(r) and HklS(r) are the solid spherical harmonic basis functions and αkli and βkli are the corresponding coefficients that are specific to the MRI system. Equation 3 clearly indicates that the actual magnetic field produced by a gradient coil consists of the sum of the intended uniform field gradient (G·r) and extra terms of higher order (E(r)). Note that the extra terms always include gradients orthogonal to the desired gradient direction.
Using the gradient calibration settings specific for an individual MR unit provided by the manufacturer, the spatial distribution of the individual gradient error functions can be calculated for each slice and used for retrospective correction of image warping and diffusion maps. To this end, equation 21 can be evaluated either numerically or analytically. Since in this formulation r=[x y z]T is the true location, the geometric distortions in the images have to be unwarped first before the diffusion-encoding correction can be applied. However, the gradient non-uniformity causes an RF pulse to excite spins that are not within a flat plane, which is also known as ‘potato chipping’. A complete correction would therefore require determining the displacement of each voxel from the desired slice and to calculate the gradient coil tensor at the correct location z′ on the ‘potato-chipped’ slice. This can be done with multi-slice or 3D data. An alternative approach which is not used here is to define the basis functions in terms of fully warped coordinates. However, the present generalized correction method can employ any other method that accurately models the spatial field distribution Bz(r) generated by the gradient coils.
To consider gradient non-linearity in the iterative reconstruction k has simply replaced by the augmented k′, i.e. the encoding the voxel truly “sees”.
Due to its simplicity the combined iterative reconstruction is very attractive. The entire recon can be even performed without knowledge of the sampling density for pre-correction. The convergence is slightly slower, though. Further improvement can be made by preconditioning or Total Variation reconstruction.
Consider now multi-shot sequences which are useful for high-resolution diffusion-weighted imaging (DWI). A common problem encountered in multi-shot image acquisition is non-linear phase variation from shot to shot. This phase error is usually caused by patient motion during diffusion encoding periods. Typically this error is corrected for by subtracting a low resolution phase map from each shot. This phase map can be obtained either from an extra navigate image or from the k-space data of a self-navigated trajectory (e.g. variable density spirals). This simple algorithm can effectively remove the phase error if image reconstruction can be separated from phase correction. However, in many cases, these two processes cannot be separated; thus cannot be performed sequentially. For example, when each shot undersamples the k-space, effect of aliasing causes the phase error at one location appearing at other locations. The resulting non-localized phase error can no longer be corrected through a simple phase subtraction. Here, we present a method that performs phase correction simultaneously with image reconstruction. This novel method combines the conjugate gradient (CG) method with a least-square estimation of the image to be reconstructed. Successful phase correction is demonstrated for multi-shot DWI with self-navigated interleaved spirals (SNAILS).
Motion during diffusion encoding periods introduces phase errors that cause serious image degradation in multi-shot DWI. This extra phase can be incorporated into the image encoding function if treated as an additive perturbation term. Written in a matrix format, data acquired with a multi-shot sequence can be expressed as,
m=Ev (24)
Here, m is the k-space data stored in a column vector; v is the image space data stored in the same fashion; and E is the encoding matrix. The size of E is Nk×N2, where Nk the total number of k-space sampling points and N is the image size. For the n-th shot, the elements of matrix E are
E(k,n),ρ=exp(−ikk,nrρ)pn(rρ) (25)
Here, kk,n is the k-th sampling point of the n-th shot, and rρis the p-th pixel of an image.
Two cases need to be considered in order to solve Eq. (24). First, if each shot samples k-space at or above the Nyquist rate (e.g. PROPELLER), then pn(rρ)m(rρ) can be easily computed for each shot separately using inverse Fourier transform. To take advantage of the fast Fourier transform (FFT) algorithm, a gridding procedure may be necessary if the sampling trajectory is non-Cartesian. Following an inverse FFT, the phase error pn(rρ) can be simply removed through a complex-number division. However, small residual artifacts will remain in the image because of the image domain convolution resulting from segmenting the k-space during data acquisition.
Second if each shot samples the k-space below the Nyquist rate (e.g. SNAILS), then pn(rρ)m(rρ) cannot be computed from each shot without suffering severe aliasing artifacts. Consequently, to reduce these artifacts, the matrix E that contains all the phase information has to be invested for both cases. This inversion can be effectively performed using a least-square estimation:
v=(EHE)−1EHm (26)
Typically E is a very large matrix. As a result, it is numerically impractical to estimate the image through direct matrix inversion and multiplication. However, this equation can be solved iteratively. For the efficiency of numerical evaluation, Eq. (26) can be rewritten as,
(EHE)v=EHm (27)
Various techniques have been developed to solve large systems of linear equations. Here, we apply the conjugate-gradient (CG) method by exploring the similarity between Eq (27) and the equation of sensitivity encoding (SENSE). We observe that Eq. (27) is equivalent to the SENSE reconstruction problem when pn(rρ) is treated as a pseudo “coil sensitivity” that varies from shot to shot.
Therefore, in the second case, to reconstruct an image from undersampled multi-shot diffusion data, we first estimate a low resolution phase map pn(rρ) for each shot. The right hand side of Eq. (27) is initially computed by directly removing the phase from each shot prior to summing all the data. During each iteration of CG, (EHE)m is calculated by using a combination of forward and backward gridding when the k-space trajectory is non-Cartesian (
In vivo diffusion-weighted images were acquired using SNAILS (3) on a GE Signal 1.5 T whole-body system. The parameters were: number of interleaf=28, FOV=22 cm, TR=2.5 s, TE=67 ms, and b=800 s/mm2. Six diffusion-encoding directions were applied to measure the diffusion tensor. Low resolution phase maps were estimated using the center k-space data from each interleaf.
We have established a mathematical framework for simultaneous image reconstruction and phase correction for multi-shot DWI. The image is computed iteratively using the CG method. In vivo experiments demonstrate that the signal cancellation caused by phase error can be successfully recovered using this algorithm (
The problem of phase correction for multi-shot DWI was previously treated by Miller et al., MRM 2003; 50: 343-353. In their approach, Eq. (26) was simplified by assuming that (EHE)−1 was an identity matrix which was equivalent to treating the data as fully sampled. The method proposed in this abstract not only corrects for phase error of fully sampled data (e.g. PROPELLER) but also corrects for undersampled data (e.g. SNAILS). If each shot samples k-space at or above the Nyquist rate, this method reduces to the straightforward phase subtraction method. Such kind of sampling strategy and phase correction scheme are seen in PROPELLER DWI. However, if each shot undersamples k-space, the simple phase subtraction method cannot completely remove the phase error. In this case, Eq. (26) needs to be solved.
Iterative Reconstruction—Typically E is a very large matrix. As a result, it is numerically impractical to estimate the image through direct matrix inversion and multiplication. An approximate solution of m can be found by assuming that (EHE)−1 is an identity matrix, which yields the refocusing reconstruction proposed by Miller and Pauly. This approximation is also equivalent to the direct phase subtraction method employed in PROPELLER DWI.
The accuracy of the approximation method depends on the k-space sampling pattern and the amount of phase error. Nevertheless, given a k-space trajectory there is no guarantee that (EHE)−1 will be an identity matrix, because the motion-induced phase error fluctuates from shot to shot and from experiment to experiment. Since it is impractical to calculate (EHE)−1 via direct matrix inversion, Eq. (2) should be solved iteratively. To improve the efficiency of the numerical evaluation, Eq. (2) can be rewritten as
(EHE)v=EHm. (28)
Various techniques have been developed to solve large systems of linear equations. For this approach, we have applied the conjugate-gradient (CG) method. Notice that Eq. [8] is equivalent to the iterative SENSE reconstruction problem when the phase error prp) is treated as a pseudo “coil sensitivity” that varies from shot to shot; pξ(rρ) is given by
pξ(rρ)=eiφ
When the k-space is non-uniformly sampled, a sampling density correction and intensity correction can be included in the computation as techniques for matrix preconditioning with the goal to reduce the number of iterations required for the CG method. The sampling density correction is performed by inserting a density matrix D in Eq. [8]:
(EHDE)v=EHD m. (30)
Here, D represents a diagonal matrix with entries equal to the inverse of the relative density of the k-space sampling pattern at position kkξ. Intensity correction is not necessary for single-coil acquisition, but will be necessary for multi-coil acquisition as discussed in the following section.
To reconstruct an image, the right hand side of Eq. (30) is initially computed by directly removing the phase from each shot prior to summing all the data. During each iteration of CG. (EHDE)v is calculated using a combination of gridding and inverse gridding (
Phase Correction and SENSE Reconstruction—By treating the phase error as a modulating function that is superimposed on the conventional image encoding function generated by the imaging gradient waveforms, we find that the mathematical formulation of the phase correction problem as shown in Eq. (28) is very similar to the SENSE reconstruction problem for arbitrary k-space acquisitions. Therefore, under the umbrella of the CG algorithm a combined reconstruction can be found for both phase-perturbed and undersampled k-space data. Specifically, for each shot or interleaf the data acquired by each coil can be considered to be encoded using a dynamic composite sensitivity profile that is the product of the coil sensitivity and the phase error. Denote sξγ the composite sensitivity profile of the γ-th coil during the ξ-th shot, and cγ the coil sensitivity profile of the γ-th coil. The composite sensitivity sξ,γ can be written as
sξ,γ(rρ)=cγeiφ
where φξγ is the phase error of the γ-th coil during the ξ-th shot in the image domain. With the definition of a composite sensitivity profile, the encoding matrix E can be defined accordingly. Specifically, each row of the matrix E contains the encoding functions for one k-space sampling point of one shot and for one coil, i.e.,
E(k,ξ,γ),ρ=e−ik
Therefore, the size of E is (nknξnγ)×N2. where nγ is number of coils. Notice that if the complex exponential in Eq. (31) is zero, Eq. (31) would yield the generalized SENSE problem.
Determination of Phase Errors—For single-coil acquisition, image reconstruction for multi-shot DWI requires an accurate Knowledge of the motion-induced phase error. Because this phase error is usually random for in vivo studies, it has to be updated in real time for each shot. One method of measuring this phase error is to acquire a low-resolution navigator image separately prior to the diffusion-weighted image acquisition. Another method is to reconstruct a low-resolution phase map using the center portion of a self-navigated k-space readout trajectory. In this study, self-navigated trajectories, including both PROPELLER and SNAILS, are utilized because of their demonstrated ability to acquire high-resolution and high-quality diffusion-weighted images. However, the specific technique for determining phase errors φξ,γ(rρ) is not critical for the phase correction algorithm introduced in this report.
Examples of PROPELLER and SNAILS trajectories are illustrated in
For multi-coil acquisition, not only the motion-induced phase error, but also the coil sensitivity (i) is required to reconstruct an image. If the multi-shot readout trajectory is self-navigated, such as PROPELLER and SNAILS, then both the motion-induced phase error and the coil sensitivity profile can be measured simultaneously using the center portion of the k-space data. In this case, the low-resolution image reconstructed from those data serves one as an estimate of the composite sensitivity profile.
Materials and Methods—The iterative CG phase correction algorithm was applied to both simulated and in vivo data. Simulations were performed on PROPELLER and SNAILS readout trajectories, whereas in vivo data were acquired using SNAILS only due to the lack of capabilities to store raw data from PROPELLER. Multi-shot SNAILS DWI was also demonstrated in combination with parallel imaging using both simulated and in vivo data.
Simulations—For single-coil simulation, artificial MRI data were created from a zero-padded (up to 4×) resolution phantom scan (signal to noise ratio (SNR)=70) by using an inverse FFT and regridding along the desired k-space trajectories (e.g. PROPELLER blades, variable-density spirals). The gridding kernel was a two-dimensional three-term cosine window function with a window width of 4 pixels. The motion effect during diffusion encoding periods was simulated by adding a linear phase term in the phantom image prior to the FFT. This phase term was randomly generated for each shot so that it corresponds to a k-space shift uniformly distributed in the range of ±6 pixels (of a 256×256 image) in each dimension.
For multi-coil simulation, an 8-channel head coil (MRI Devices Corporation, Pewaukee, Wis.) was used. The sensitivity of each coil element was measured using a water phantom; the sensitivity was computed by dividing each coil image with the body-coil image (xii). To simulate the data acquisition for each coil during each shot, the resolution phantom image was multiplied by each coil's sensitivity. To simulate the motion-induced phase error, a random linear phase term was added to the image. The k-space data were then computed following the same procedure described in the single-coil simulation.
Single-coil simulation was performed using both PROPELLER and SNAILS trajectory, while, for simplicity, multi-coil simulation was only performed using SNAILS. The PROPELLER trajectory consisted of 28 blades. Each blade, which was uniformly sampled at the Nyquist rate, had 14 phase encoding lines. The SNAILS trajectory contained 23 interleaves of variable density (VD) spirals which had a pitch factor α of 4. Each spiral consisted of 3736 sampling points. To reconstruct the image from the simulated k-space data, the true phase error and coil sensitivity profile instead of estimated phase error and sensitivity were used in the reconstruction for the purpose of illustrating the accuracy of the CG phase correction algorithm itself.
Experiments—For single-coil acquisition, in vivo diffusion-weighted images were acquired using SNAILS on a 1.5T whole-body system (GE Signal, GE Healthcare, Waukesha, Wis.) equipped with a maximum gradient of 50 mT/m and a slew rate of 150 mT/m/s. A twice-refocused spin echo sequence was used to suppress gradient-induced eddy currents. The scan parameters were: FOV=22 cm, TR=2.5 s, TE=67 ms, bandwidth (BW)=125 kHz, and b=800 s/mm2. The duration of the four diffusion gradients was: δ1=δ3=8.4 ms and δ2=δ4=12.3 ms. A total of 20 axial slices were prescribed with a 6.5 mm slice thickness and a zero gap in between. With an acquisition matrix size of 256×256, an in-plane resolution of 0.86×0.86 mm2 was achieved. The VD spiral readout trajectory was the same as that used in the simulations (i.e. 23 interleaves, 3736 samples/interleaf, α=4), which resulted in a readout time of 13.5 ms for each interleaf. Of the 3736 samples, the first 150 samples (that formed the center portion of the k-space) also contributed to the determination of the composite sensitivity profiles.
For multi-coil acquisition, in vivo diffusion-weighted images were acquired using the 8-channel head coil (MRI Devices Corporation, Pewaukee, Wis.). All sequence parameters were kept the same as for the single-coil acquisition. The reduction factor ranges from R=1 to 3.
Six diffusion-encoding directions were applied to measure the diffusion tensor. The six directions are (1 1 0)T, (0 1 1)T, (1 0 1)T, (1 −1 0)T, (0 1 −1)T, and (1 0 −1)T. The scan time for acquiring six diffusion-weighted images plus an unweighted image was 6.7 minutes for coverage of the entire brain. For single-coil acquisition, this scan was repeated five times; for multi-coil acquisition, it was repeated twice.
All image reconstructions were performed using Matlab (Version 7, Release 14, The MathWorks, Inc.), running on a Windows PC equipped with a 3.06 GHz Pentium 4 CPU and 1.5 GB RAM.
CG Phase Correction—
Phase Correction and SENSE Reconstruction—Multi-coil SNAILS data with random linear phase error were simulated using a SENSE reduction factor (xii) ranging from 1 to 3. The reconstructed images are displayed in
From the acquired in vivo multi-coil data, images are reconstructed with different SENSE reduction factors. To obtain an image with a reduction factor of 2, data from a total of 11 interleaves are used by skipping every other interleaf. To achieve a reduction factor of 3, 8 interleaves are used. The reconstructed images are shown in
We have established a mathematical framework of simultaneous phase correction and parallel imaging (SENSE) reconstruction for multi-shot DWI using arbitrary k-space trajectories. By treating the phase error as an additional modulation factor to the existing gradient and sensitivity encoding function, we can express the k-space data as a linear combination of the image-domain data. The resultant set of linear equations can be solved iteratively using the conjugate gradient method. Both simulations and in vivo experiments have demonstrated that the CG phase correction method can effectively reduce artifacts caused by motion-induced phase errors, and reconstruct diffusion-weighted images acquired with multiple coils using R=1 to 3.
CG Phase Correction—The problem of phase correction for multi-shot DWI was previously treated as a linear estimation problem by Miller and Pauly, supra. In that approach, Eq. (24) was simplified by approximating (EHE)−1 as an identity matrix, which resulted in the direct phase subtraction method. Miller and Pauly also showed that the accuracy of this approximation could be improved by averaging the data over multiple acquisitions, talking advantage of destructive interference of the residual aliasing. Using this direct phase subtraction method, some high-quality diffusion-weighted images were previously acquired with PROPELLER and navigated interleaved EPI techniques.
However, the direct phase subtraction method does not completely correct the effect of motion-induced phase errors due to image blurring or aliasing, as illustrated by our computer simulations. Specifically, for a segmented acquisition, such as PROPELLER, segmenting k-space results in a convolution in the image space, which causes the phase error of one pixel to appear in adjacent pixels. Consequently, the straightforward phase subtraction results in an incomplete phase correction and a relatively high background noise (
The CG phase correction method is a generalized phase correction technique for multi-shot DWI. It does not rely on specific readout trajectories, though certain trajectories might take advantages of this method more easily than others. Despite the fact that different k-space strategies result in different sensitivities to motion artifacts, the motion correction problem can be formulated in the same way and solved using the CG method. For example, the motion artifacts of PROPELLER DWI are mostly localized, which means that the distortion at one location is mainly caused by the interference of its neighboring structures (
Accurate phase correction and SENSE reconstruction relies on an accurate knowledge of phase errors and the coil sensitivity profiles. Because the phase errors change from shot to shot, it has to be measured for each shot. If the subject motion is small, the low resolution phase error in the spatial domain can be estimated accurately using the center k-space data. However, if the phase error contains rapidly varying components, a low-resolution phase map will not be sufficient. As a result, there will be residual artifacts in the image even after phase correction. Other techniques that have been developed to alleviate the problem of inaccurate sensitivity profiles can be also applied here.
Multi-Shot SENSE DWI—Besides using multi-shot sequences, another approach for high resolution DWI is to combine parallel imaging with single-shot EPI. However, there have been no reported attempts to combine parallel imaging techniques with multi-shot DWI. One difficulty for multi-shot SENSE DWI is performing phase correction on data modulated by both coil sensitivity and motion-induced phase error. For single-shot acquisitions, the motion-induced phase error p(rρ) is a common factor for all coils, thus can be absorbed into the vector m and neglected in the image reconstruction, if one only cares about the magnitude of the image. However, for multi-shot acquisitions, even if the sensitivity profile of each coil might be stable during the experiment, motion-induced phase errors vary from shot to shot and cannot be ignored.
By treating the phase error as an additional image encoding function, we find that the phase error and the coil sensitivity can be combined to form a composite sensitivity profile for each receiver coil element. With this treatment. Phase correction and SENSE reconstruction can be formulated in the same set of linear equations, which can be solved iteratively with the CG method. The feasibility of simultaneous phase correction and SENSE reconstruction has been demonstrated with both simulation (
The proposed reconstruction method for multi-shot SENSE DWI is computationally demanding, especially when using the forward and backward gridding. The forward and backward gridding procedure has to be performed independently for each interleaf and for each coil element. There are typically many interleaves for high resolution acquisitions: for example, SNAILS requires 23 interleaves to fully sample an image of 256×256 pixels. With increasing number of receiver coils the reconstruction becomes increasingly demanding. When an 8-channel receiver coil is used, for example, this gridding procedure has to be performed 184 times for each CG iteration (
One drawback of multi-shot DWT sequences has been the increase of scan time. The scan-time penalty is a result of the fundamental tradeoff between resolution and SNR. The proposed multi-shot SENSE technique facilitates the application of parallel imaging in multi-shot DWI and therefore reduces the minimum scan time required. As demonstrated by the in vivo experiments, using a reduction factor of 2, a whole-brain diffusion tensor imaging (DTI) study with a resolution of 256×256 can be performed in less than 7 minutes. The resultant images exhibit good quality. Another benefit of parallel imaging with multi-shot DWI, for example SNAILS, is the capability of reducing off-resonance blurring. With a reduction factor of 2, the readout time can be shortened by prescribing more interleaves; therefore image blurring caused by off-resonant spinis can be reduced.
By treating the motion-induced phase error as an additional image encoding function, the phase correction problem can be formulated as a set of linear equations. This set of equations can be solved iteratively using the CG method. Simulations and in vivo studies have shown that this method can significantly reduce the residual artifacts resulting from the direct phase subtraction method. Furthermore, because its similarity to sensitivity encoding, the motion-induced phase error can be incorporated into the SENSE reconstruction by defining a dynamic composite sensitivity profile. This simultaneous phase correction and SENSE reconstruction allows for high resolution multi-shot SENSE DWI.
While the invention has been described with reference to specific embodiments, the description is illustrative of the invention and not limiting the invention. Various applications and modifications may occur to those skilled in the art without departing from the true spirit and scope of the invention as defined by the appended claims.
Patent | Priority | Assignee | Title |
10115212, | Nov 07 2016 | UIH AMERICA, INC ; SHANGHAI UNITED IMAGING HEALTHCARE CO , LTD | Image reconstruction system and method in magnetic resonance imaging |
10302726, | Jul 03 2014 | KONINKLIJKE PHILIPS N V | Image reconstruction for MRI using multiplexed sensitivity encoding |
10373351, | Nov 07 2016 | SHANGHAI UNITED IMAGING HEALTHCARE CO., LTD | Image reconstruction system and method in magnetic resonance imaging |
10429471, | Oct 21 2015 | KONINKLIJKE PHILIPS N V | MRI with variable density sampling |
10551458, | Jun 29 2017 | General Electric Company | Method and systems for iteratively reconstructing multi-shot, multi-acquisition MRI data |
10600215, | Nov 07 2016 | SHANGHAI UNITED IMAGING HEALTHCARE CO., LTD. | Image reconstruction system and method in magnetic resonance imaging |
10823799, | Mar 14 2017 | SIEMENS HEALTHINEERS AG | Method and apparatus for accelerated generation of a series of magnetic resonance images with simultaneous multislice imaging |
11085988, | Mar 20 2019 | The Board of Trustees of the Leland Stanford Junior University | Method for estimating systematic imperfections in medical imaging systems with deep learning |
11112475, | Jun 09 2017 | KONINKLIJKE PHILIPS N V | Parallel multi-slice MR imaging |
7622924, | Jun 12 2007 | General Electric Company | Method and apparatus for k-space and hybrid-space based image reconstruction for parallel imaging and artifact correction |
7728588, | Sep 13 2006 | SIEMENS HEALTHINEERS AG | Magnetic resonance imaging data acquisition sequence and apparatus |
7786728, | Nov 02 2007 | Toshiba Medical Systems Corporation | Magnetic resonance imaging apparatus and magnetic resonance imaging method |
7853060, | Apr 13 2006 | SIEMENS HEALTHINEERS AG | Method and MR system for generating MR images |
7860291, | Jun 12 2007 | General Electric Co. | Method and apparatus for correcting motion in multi-shot diffusion-weighted magnetic resonance imaging |
7863898, | Mar 13 2008 | Siemens Aktiengesellschaft | Method and magnetic resonance system to determine the phase position of a magnetization |
7902825, | May 19 2008 | The Board of Trustees of the Leland Stanford Junior University | Motion corrected tensor magnetic resonance imaging |
7994784, | Jul 20 2007 | AUGUSTA UNIVERSITY RESEARCH INSTITUTE, INC | Systems and methods for rescaling image intensities with diffusion MRI to synthesize complex diffusive geometries |
8098068, | Jul 20 2007 | AUGUSTA UNIVERSITY RESEARCH INSTITUTE, INC | Systems and methods for parametric mapping for correction of noise-based systematic bias of DTI metrics, using a DTI mapping phantom |
8134363, | Jul 20 2007 | AUGUSTA UNIVERSITY RESEARCH INSTITUTE, INC | Test object for use with diffusion MRI and system and method of synthesizing complex diffusive geometries using novel gradient directions |
8154293, | Jul 31 2009 | General Electric Company | Apparatus and method for acquisition and reconstruction of non-CPMG propeller sequences |
8169216, | Jul 20 2007 | AUGUSTA UNIVERSITY RESEARCH INSTITUTE, INC | System and method for synthesizing crossing ADC distributions via reassembly of multiple k-spaces |
8248071, | Feb 12 2008 | Research Foundation of State University of New York, The | Methods of using combined forward and backward sampling of nuclear magnetic resonance time domain for measurement of secondary phase shifts, detection of absorption mode signals devoid of dispersive components, and/or optimization of nuclear magnetic resonance experiments |
8314614, | May 03 2007 | Koninklijke Philips Electronics N V | PROPELLER MRI with phase correction |
8334695, | Mar 25 2010 | Board of Regents of the University of Texas System | Method and apparatus for improving the quality of MR images sensitized to molecular diffusion |
8447089, | Apr 27 2010 | Systems and methods for susceptibility tensor imaging | |
8482280, | Jan 23 2009 | DIGNITY HEALTH | System and method for propeller magnetic resonance imaging with non-parallel bladelets |
9285449, | Jun 15 2011 | Systems and methods for imaging and quantifying tissue magnetism with magnetic resonance imaging | |
9329252, | Aug 30 2011 | The Board of Trustees of the Leland Stanford Junior University; BOARD OF TRUSTEES OF THE LELAND STANFORD JUNIOR UNIVERSITY, THE | Apparatus for real-time phase correction for diffusion-weighted magnetic resonance imaging using adaptive RF pulses |
9383423, | Apr 27 2010 | Systems and methods for susceptibility tensor imaging | |
9897678, | Apr 19 2013 | General Electric Company | Magnetic resonance imaging data correction methods and systems |
Patent | Priority | Assignee | Title |
6853191, | Dec 10 2003 | The Board of Trustees of the Leland Stanford Junior University | Method of removing dynamic nonlinear phase errors from MRI data |
20070182411, |
Executed on | Assignor | Assignee | Conveyance | Frame | Reel | Doc |
Jan 26 2006 | BAMMER, ROLAND | BOARD OF TRUSTEES OF THE LELAND STANFORD JUNIOR UNIVERSITY, THE | ASSIGNMENT OF ASSIGNORS INTEREST SEE DOCUMENT FOR DETAILS | 017573 | /0790 | |
Jan 26 2006 | LIU, CHUNLEI | BOARD OF TRUSTEES OF THE LELAND STANFORD JUNIOR UNIVERSITY, THE | ASSIGNMENT OF ASSIGNORS INTEREST SEE DOCUMENT FOR DETAILS | 017573 | /0790 | |
Feb 06 2006 | The Board of Trustees of the Leland Stanford Junior University | (assignment on the face of the patent) | / | |||
Dec 08 2006 | Stanford University | NATIONAL INSTITUTES OF HEALTH NIH , U S DEPT OF HEALTH AND HUMAN SERVICES DHHS , U S GOVERNMENT | CONFIRMATORY LICENSE SEE DOCUMENT FOR DETAILS | 024758 | /0638 |
Date | Maintenance Fee Events |
Feb 06 2012 | M1551: Payment of Maintenance Fee, 4th Year, Large Entity. |
Feb 05 2016 | M1552: Payment of Maintenance Fee, 8th Year, Large Entity. |
Jan 30 2020 | M1553: Payment of Maintenance Fee, 12th Year, Large Entity. |
Date | Maintenance Schedule |
Aug 05 2011 | 4 years fee payment window open |
Feb 05 2012 | 6 months grace period start (w surcharge) |
Aug 05 2012 | patent expiry (for year 4) |
Aug 05 2014 | 2 years to revive unintentionally abandoned end. (for year 4) |
Aug 05 2015 | 8 years fee payment window open |
Feb 05 2016 | 6 months grace period start (w surcharge) |
Aug 05 2016 | patent expiry (for year 8) |
Aug 05 2018 | 2 years to revive unintentionally abandoned end. (for year 8) |
Aug 05 2019 | 12 years fee payment window open |
Feb 05 2020 | 6 months grace period start (w surcharge) |
Aug 05 2020 | patent expiry (for year 12) |
Aug 05 2022 | 2 years to revive unintentionally abandoned end. (for year 12) |